Compressed Decentralized Momentum Stochastic Gradient Methods for Nonconvex Optimization

TL;DR

This paper introduces two novel compressed decentralized stochastic gradient algorithms with momentum for nonconvex optimization, achieving optimal convergence and superior empirical results on neural networks.

Contribution

It presents the first decentralized adaptive method with compression and a heavy-ball method addressing data heterogeneity, both with proven optimal convergence.

Findings

Achieves optimal convergence rates for both algorithms.

Demonstrates linear speedup and topology-independent parameters.

Outperforms state-of-the-art methods on neural network training.

Abstract

In this paper, we design two compressed decentralized algorithms for solving nonconvex stochastic optimization under two different scenarios. Both algorithms adopt a momentum technique to achieve fast convergence and a message-compression technique to save communication costs. Though momentum acceleration and compressed communication have been used in literature, it is highly nontrivial to theoretically prove the effectiveness of their composition in a decentralized algorithm that can maintain the benefits of both sides, because of the need to simultaneously control the consensus error, the compression error, and the bias from the momentum gradient. For the scenario where gradients are bounded, our proposal is a compressed decentralized adaptive method. To the best of our knowledge, this is the first decentralized adaptive stochastic gradient method with compressed communication. For the scenario of data heterogeneity without bounded gradients, our proposal is a compressed decentralized heavy-ball method, which applies a gradient tracking technique to address the challenge of data heterogeneity. Notably, both methods achieve an optimal convergence rate, and they can achieve linear speed up and adopt topology-independent algorithmic parameters within a certain regime of the user-specified error tolerance. Superior empirical performance is observed over state-of-the-art methods on training deep neural networks (DNNs) and Transformers.